1/*
2 * pseries CPU Hotplug infrastructure.
3 *
4 * Split out from arch/powerpc/platforms/pseries/setup.c
5 *  arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
6 *
7 * Peter Bergner, IBM	March 2001.
8 * Copyright (C) 2001 IBM.
9 * Dave Engebretsen, Peter Bergner, and
10 * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11 * Plus various changes from other IBM teams...
12 *
13 * Copyright (C) 2006 Michael Ellerman, IBM Corporation
14 *
15 *      This program is free software; you can redistribute it and/or
16 *      modify it under the terms of the GNU General Public License
17 *      as published by the Free Software Foundation; either version
18 *      2 of the License, or (at your option) any later version.
19 */
20
21#include <linux/kernel.h>
22#include <linux/interrupt.h>
23#include <linux/delay.h>
24#include <linux/sched.h>	/* for idle_task_exit */
25#include <linux/cpu.h>
26#include <linux/of.h>
27#include <asm/prom.h>
28#include <asm/rtas.h>
29#include <asm/firmware.h>
30#include <asm/machdep.h>
31#include <asm/vdso_datapage.h>
32#include <asm/xics.h>
33#include <asm/plpar_wrappers.h>
34
35#include "offline_states.h"
36
37/* This version can't take the spinlock, because it never returns */
38static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
39
40static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
41							CPU_STATE_OFFLINE;
42static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
43
44static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
45
46static int cede_offline_enabled __read_mostly = 1;
47
48/*
49 * Enable/disable cede_offline when available.
50 */
51static int __init setup_cede_offline(char *str)
52{
53	if (!strcmp(str, "off"))
54		cede_offline_enabled = 0;
55	else if (!strcmp(str, "on"))
56		cede_offline_enabled = 1;
57	else
58		return 0;
59	return 1;
60}
61
62__setup("cede_offline=", setup_cede_offline);
63
64enum cpu_state_vals get_cpu_current_state(int cpu)
65{
66	return per_cpu(current_state, cpu);
67}
68
69void set_cpu_current_state(int cpu, enum cpu_state_vals state)
70{
71	per_cpu(current_state, cpu) = state;
72}
73
74enum cpu_state_vals get_preferred_offline_state(int cpu)
75{
76	return per_cpu(preferred_offline_state, cpu);
77}
78
79void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
80{
81	per_cpu(preferred_offline_state, cpu) = state;
82}
83
84void set_default_offline_state(int cpu)
85{
86	per_cpu(preferred_offline_state, cpu) = default_offline_state;
87}
88
89static void rtas_stop_self(void)
90{
91	static struct rtas_args args = {
92		.nargs = 0,
93		.nret = cpu_to_be32(1),
94		.rets = &args.args[0],
95	};
96
97	args.token = cpu_to_be32(rtas_stop_self_token);
98
99	local_irq_disable();
100
101	BUG_ON(rtas_stop_self_token == RTAS_UNKNOWN_SERVICE);
102
103	printk("cpu %u (hwid %u) Ready to die...\n",
104	       smp_processor_id(), hard_smp_processor_id());
105	enter_rtas(__pa(&args));
106
107	panic("Alas, I survived.\n");
108}
109
110static void pseries_mach_cpu_die(void)
111{
112	unsigned int cpu = smp_processor_id();
113	unsigned int hwcpu = hard_smp_processor_id();
114	u8 cede_latency_hint = 0;
115
116	local_irq_disable();
117	idle_task_exit();
118	xics_teardown_cpu();
119
120	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
121		set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
122		if (ppc_md.suspend_disable_cpu)
123			ppc_md.suspend_disable_cpu();
124
125		cede_latency_hint = 2;
126
127		get_lppaca()->idle = 1;
128		if (!lppaca_shared_proc(get_lppaca()))
129			get_lppaca()->donate_dedicated_cpu = 1;
130
131		while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
132			while (!prep_irq_for_idle()) {
133				local_irq_enable();
134				local_irq_disable();
135			}
136
137			extended_cede_processor(cede_latency_hint);
138		}
139
140		local_irq_disable();
141
142		if (!lppaca_shared_proc(get_lppaca()))
143			get_lppaca()->donate_dedicated_cpu = 0;
144		get_lppaca()->idle = 0;
145
146		if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
147			unregister_slb_shadow(hwcpu);
148
149			hard_irq_disable();
150			/*
151			 * Call to start_secondary_resume() will not return.
152			 * Kernel stack will be reset and start_secondary()
153			 * will be called to continue the online operation.
154			 */
155			start_secondary_resume();
156		}
157	}
158
159	/* Requested state is CPU_STATE_OFFLINE at this point */
160	WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
161
162	set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
163	unregister_slb_shadow(hwcpu);
164	rtas_stop_self();
165
166	/* Should never get here... */
167	BUG();
168	for(;;);
169}
170
171static int pseries_cpu_disable(void)
172{
173	int cpu = smp_processor_id();
174
175	set_cpu_online(cpu, false);
176	vdso_data->processorCount--;
177
178	/*fix boot_cpuid here*/
179	if (cpu == boot_cpuid)
180		boot_cpuid = cpumask_any(cpu_online_mask);
181
182	/* FIXME: abstract this to not be platform specific later on */
183	xics_migrate_irqs_away();
184	return 0;
185}
186
187/*
188 * pseries_cpu_die: Wait for the cpu to die.
189 * @cpu: logical processor id of the CPU whose death we're awaiting.
190 *
191 * This function is called from the context of the thread which is performing
192 * the cpu-offline. Here we wait for long enough to allow the cpu in question
193 * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
194 * notifications.
195 *
196 * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
197 * self-destruct.
198 */
199static void pseries_cpu_die(unsigned int cpu)
200{
201	int tries;
202	int cpu_status = 1;
203	unsigned int pcpu = get_hard_smp_processor_id(cpu);
204
205	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
206		cpu_status = 1;
207		for (tries = 0; tries < 5000; tries++) {
208			if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
209				cpu_status = 0;
210				break;
211			}
212			msleep(1);
213		}
214	} else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
215
216		for (tries = 0; tries < 25; tries++) {
217			cpu_status = smp_query_cpu_stopped(pcpu);
218			if (cpu_status == QCSS_STOPPED ||
219			    cpu_status == QCSS_HARDWARE_ERROR)
220				break;
221			cpu_relax();
222		}
223	}
224
225	if (cpu_status != 0) {
226		printk("Querying DEAD? cpu %i (%i) shows %i\n",
227		       cpu, pcpu, cpu_status);
228	}
229
230	/* Isolation and deallocation are definitely done by
231	 * drslot_chrp_cpu.  If they were not they would be
232	 * done here.  Change isolate state to Isolate and
233	 * change allocation-state to Unusable.
234	 */
235	paca[cpu].cpu_start = 0;
236}
237
238/*
239 * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
240 * here is that a cpu device node may represent up to two logical cpus
241 * in the SMT case.  We must honor the assumption in other code that
242 * the logical ids for sibling SMT threads x and y are adjacent, such
243 * that x^1 == y and y^1 == x.
244 */
245static int pseries_add_processor(struct device_node *np)
246{
247	unsigned int cpu;
248	cpumask_var_t candidate_mask, tmp;
249	int err = -ENOSPC, len, nthreads, i;
250	const __be32 *intserv;
251
252	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
253	if (!intserv)
254		return 0;
255
256	zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
257	zalloc_cpumask_var(&tmp, GFP_KERNEL);
258
259	nthreads = len / sizeof(u32);
260	for (i = 0; i < nthreads; i++)
261		cpumask_set_cpu(i, tmp);
262
263	cpu_maps_update_begin();
264
265	BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
266
267	/* Get a bitmap of unoccupied slots. */
268	cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
269	if (cpumask_empty(candidate_mask)) {
270		/* If we get here, it most likely means that NR_CPUS is
271		 * less than the partition's max processors setting.
272		 */
273		printk(KERN_ERR "Cannot add cpu %s; this system configuration"
274		       " supports %d logical cpus.\n", np->full_name,
275		       num_possible_cpus());
276		goto out_unlock;
277	}
278
279	while (!cpumask_empty(tmp))
280		if (cpumask_subset(tmp, candidate_mask))
281			/* Found a range where we can insert the new cpu(s) */
282			break;
283		else
284			cpumask_shift_left(tmp, tmp, nthreads);
285
286	if (cpumask_empty(tmp)) {
287		printk(KERN_ERR "Unable to find space in cpu_present_mask for"
288		       " processor %s with %d thread(s)\n", np->name,
289		       nthreads);
290		goto out_unlock;
291	}
292
293	for_each_cpu(cpu, tmp) {
294		BUG_ON(cpu_present(cpu));
295		set_cpu_present(cpu, true);
296		set_hard_smp_processor_id(cpu, be32_to_cpu(*intserv++));
297	}
298	err = 0;
299out_unlock:
300	cpu_maps_update_done();
301	free_cpumask_var(candidate_mask);
302	free_cpumask_var(tmp);
303	return err;
304}
305
306/*
307 * Update the present map for a cpu node which is going away, and set
308 * the hard id in the paca(s) to -1 to be consistent with boot time
309 * convention for non-present cpus.
310 */
311static void pseries_remove_processor(struct device_node *np)
312{
313	unsigned int cpu;
314	int len, nthreads, i;
315	const __be32 *intserv;
316	u32 thread;
317
318	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
319	if (!intserv)
320		return;
321
322	nthreads = len / sizeof(u32);
323
324	cpu_maps_update_begin();
325	for (i = 0; i < nthreads; i++) {
326		thread = be32_to_cpu(intserv[i]);
327		for_each_present_cpu(cpu) {
328			if (get_hard_smp_processor_id(cpu) != thread)
329				continue;
330			BUG_ON(cpu_online(cpu));
331			set_cpu_present(cpu, false);
332			set_hard_smp_processor_id(cpu, -1);
333			break;
334		}
335		if (cpu >= nr_cpu_ids)
336			printk(KERN_WARNING "Could not find cpu to remove "
337			       "with physical id 0x%x\n", thread);
338	}
339	cpu_maps_update_done();
340}
341
342static int pseries_smp_notifier(struct notifier_block *nb,
343				unsigned long action, void *data)
344{
345	struct of_reconfig_data *rd = data;
346	int err = 0;
347
348	switch (action) {
349	case OF_RECONFIG_ATTACH_NODE:
350		err = pseries_add_processor(rd->dn);
351		break;
352	case OF_RECONFIG_DETACH_NODE:
353		pseries_remove_processor(rd->dn);
354		break;
355	}
356	return notifier_from_errno(err);
357}
358
359static struct notifier_block pseries_smp_nb = {
360	.notifier_call = pseries_smp_notifier,
361};
362
363#define MAX_CEDE_LATENCY_LEVELS		4
364#define	CEDE_LATENCY_PARAM_LENGTH	10
365#define CEDE_LATENCY_PARAM_MAX_LENGTH	\
366	(MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
367#define CEDE_LATENCY_TOKEN		45
368
369static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
370
371static int parse_cede_parameters(void)
372{
373	memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
374	return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
375			 NULL,
376			 CEDE_LATENCY_TOKEN,
377			 __pa(cede_parameters),
378			 CEDE_LATENCY_PARAM_MAX_LENGTH);
379}
380
381static int __init pseries_cpu_hotplug_init(void)
382{
383	struct device_node *np;
384	const char *typep;
385	int cpu;
386	int qcss_tok;
387
388	for_each_node_by_name(np, "interrupt-controller") {
389		typep = of_get_property(np, "compatible", NULL);
390		if (strstr(typep, "open-pic")) {
391			of_node_put(np);
392
393			printk(KERN_INFO "CPU Hotplug not supported on "
394				"systems using MPIC\n");
395			return 0;
396		}
397	}
398
399	rtas_stop_self_token = rtas_token("stop-self");
400	qcss_tok = rtas_token("query-cpu-stopped-state");
401
402	if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
403			qcss_tok == RTAS_UNKNOWN_SERVICE) {
404		printk(KERN_INFO "CPU Hotplug not supported by firmware "
405				"- disabling.\n");
406		return 0;
407	}
408
409	ppc_md.cpu_die = pseries_mach_cpu_die;
410	smp_ops->cpu_disable = pseries_cpu_disable;
411	smp_ops->cpu_die = pseries_cpu_die;
412
413	/* Processors can be added/removed only on LPAR */
414	if (firmware_has_feature(FW_FEATURE_LPAR)) {
415		of_reconfig_notifier_register(&pseries_smp_nb);
416		cpu_maps_update_begin();
417		if (cede_offline_enabled && parse_cede_parameters() == 0) {
418			default_offline_state = CPU_STATE_INACTIVE;
419			for_each_online_cpu(cpu)
420				set_default_offline_state(cpu);
421		}
422		cpu_maps_update_done();
423	}
424
425	return 0;
426}
427machine_arch_initcall(pseries, pseries_cpu_hotplug_init);
428